[computer-go] Re: Scalability

[Don Dailey <drd@xxxxxxx>]

I'm not sure whether this is semantics or not, but it is very
interesting that in chess, it's kind of the opposite.  In general, the
capture search boils down to punishing (or refuting) stupid moves
usually just played.  You often need to see if your last move was a
stupid piece hanging blunder.    Without quies, you play into tactics
where some piece is hanging at the very end.   

The key is that you are looking for something your evaluation doesn't
understand.   Even though it seems opposite to your answer,  this much
is the same.    

So maybe a good guide to what a quies search should include is any
move that helps resolve positions that your evaluation doesn't
understand on it's own.  This would encompass "responses to disturbing
moves" and "refutations of blunders."

- Don



   My simple answer to your question below is that Q-search in go are
   responses to "disturbing moves". Most of these moves are bad sente
   moves. For example meaningless peeps or attacks on keimas that cannot be
   cut.

   If you make a 1-ply search without pruning such bad moves the evaluation
   function (well at least the one I programmed) will return a high score
   for all those moves. 
   A q-search would respond with those simple local moves that repairs the
   damage and even shows a gain for the defender. 

   So Q-search in go would be more of repairing the shape of stable groups
   rather than life and death and captures.

   My program Viking does this by matching patterns to centered on each
   move played in 1ply full board look ahead. If a Q-detection pattern
   matches then it will make a local response (effectively 2-ply locally).
   It also does a 1-ply search with similar replies if necessary for all
   opponent moves. In order to detect strong attacks that has no strong
   response.

   This work very well compared to a simple 1-ply search and is faster than
   a 2-ply search, but often it breaks down in complex fighting (but which
   go program does not have that problem?)

   --
   Magnus Persson
   Center for Adaptive Behavior and Cognition
   Tel: +49-(30)-82406-350
   Cell phone: +49 163 6639868

   > -----Original Message-----
   > From: Don Dailey [mailto:drd@xxxxxxxxxxxxxxxxx]
   > 
   > >  So at least in principle traditional chess techniques work in go
   but
   > >  since you only can search 2 ply with a complex evaluation function
   all
   > >  the cool chess stuff such as null moves and transposition tables
   has no
   > >  impact.
   > 
   > Yes, it seems so.  But I'm not even sure what the term "quiescence
   > search" really means in Go.  It is really an ambigious term to
   > describe a kind of "selectivity done just before leaf nodes."


   > Are some of you guys using such a search in your GO programs?  I'm not
   > talking about the highly localized life/death searches which isn't in
   > the spirit of giving a quiet board.
   > 
   > - Don

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/